Third-Party AI Evals Don't Fix Secrecy, Argues Researcher 1a3orn
1a3orn · x · 2026-09-16
1a3orn argues that letting third-party evaluators into AI labs helps somewhat with principal-agent problems, but does nothing for science or building common knowledge. Quoting labs that call AI super dangerous while refusing to disclose pretraining data, architectures, post-training algorithms, or RL environments, he contends that with everyone "flailing in ignorance," releasing substantive public information matters more right now.
Related event: AI Labs' Double Talk on Danger Sparks Fierce Debate(6 posts)→
More from AGI Musings
- Richard Socher: There is no realistic scenario where AI wipes out all of humanity — dilipkay · 2026-09-16
- David Deutsch: AIs are harmless, but trying to enslave AGIs is fatal — inductionheads · 2026-09-16
- Researcher: talk directly to AI labs' biosecurity teams, not their CEOs' polished takes — anshulkundaje · 2026-09-16
- New preprint 'The No Body Problem' asks whether intelligence is possible without a body — AnnaCiaunica · 2026-09-16
- Defending AI consciousness research: investigating model welfare is sane, not silly — rgblong · 2026-09-16
- Researcher: AI lab employees are genuinely worried, but conspiracies are used to dismiss concerns — NathanpmYoung · 2026-09-16